The AGORÀ Briefing — Friday, October 16, 2026

Good morning, Adam. This report speaks first of all to boards of directors. In October 2026, the Financial Stability Board delivered to the G20 the final version of its 12 practices for the responsible adoption of AI. The practices are voluntary. The text asks boards and senior management to use them as a reference for strategy, technology choices, and risk management. The push comes from generative AI and agentic AI. Those are systems that carry out multi-step tasks on their own. The ABA Banking Journal reported on the final text, which was wrapped up after a public consultation and a workshop. Here is how I read it. We are in the phase that comes before mandatory rules. Banking supervision has been here before. On June 26, 1974, German authorities shut down Herstatt Bank in the middle of the trading day. That crisis gave birth to the Basel Committee on Banking Supervision. Back then, too, the principles came first. The capital rules came after.

Hello, Adam. India's child rights authority, the NCPCR, has called Google in for a meeting by the end of October. The issue is the "kiss and hug" AI apps on the Play Store. According to MediaNama, these apps create intimate videos from photos of real people, and they skip every consent check. On October 14, ThePrint reported that some of these apps were rated as suitable for children aged three and up. In New Delhi, at the Trust and Safety India Festival 2026, NCPCR secretary Sanjeev Sharma announced child safety guidelines for social media, games, and AI platforms by December. The IT (Amendment) Rules, 2026, which took effect on February 10, 2026, require platforms to label synthetic content. But MediaNama says it is still unclear whether those labeling rules cover apps that only generate the videos.

Good morning, everyone. A company called Brennan reviewed a professional services firm before that firm switched on Microsoft Copilot, iTnews reported on October 14, 2026. At that firm, one question to Copilot in everyday language was enough to pull up confidential files. The story comes from a sponsored piece published on October 14 on iTnews. Before the firm turned on Copilot in Microsoft 365, Brennan ran a full review of its systems. The firm came out strong in four areas and weak in two. Those two were information classification and data governance. SharePoint was the main document archive. According to the report, the sharing settings let Copilot show users files they technically had access to but would never have gone looking for themselves. The research firm ADAPT, quoted in the same article, says the problem is widespread. ADAPT surveyed more than 200 data and technology leaders in Australia and New Zealand in April 2026. It found that about half of the AI tools those companies are testing or using are not covered by any formal rules.

Hi, everyone. The posting was looking for an executive assistant in London, with a salary of 145,000 euros. Fast Company puts that at about 190,000 dollars. Anthropic took the posting off its website, but Fast Company kept a copy. City AM picked up the story too, and tied it to the race toward AI agents. The role supports senior leaders in sales and marketing. The duties are the classic ones. Calendars, travel, email, expense reports. The personal qualities in the posting tell you more about the job. The text asks for exceptional attention to detail, strong relationship-building skills, and the ability to anticipate the needs of different teams. The work has to be done on time and without mistakes. Social media users quoted by Fast Company pointed out the irony. A company leading the way on AI agents is hiring a person for a job that, in theory, an agent could largely handle.

Hi, Adam. The product to watch is Brand Concierge, the assistant that talks to customers on brand websites. On October 15, Adobe and NVIDIA expanded the partnership they announced in March. The products involved are Photoshop, Premiere, Substance 3D, Frame.io, and Brand Concierge. The two companies say Photoshop, Premiere, and Substance 3D run up to 2 times faster on NVIDIA's RTX Spark superchip. That figure comes from them. It applies to AI processing, editing, color, and effects on the individual creator's machine. Frame.io uses NVIDIA GPUs to upload and analyze files. Brand Concierge now runs on Nemotron, NVIDIA's open models. Work continues on Firefly Foundry and on 3D digital twins for businesses. For anyone running a company, the assistant that talks to customers matters more than faster editing on a creator's computer.

Good morning to everyone listening. In my view, the grid will soon stop holding AI back. According to a count by Cleanview, reported by NextBigFuture, in mid-2026 the United States has 2 GW of data centers that make their own power, with a plant right on campus. Cleanview estimates about 3 GW by the end of 2026, and up to 13 GW by the end of 2027. According to the report, that capacity exceeds New York's electricity demand. Meta is doing exactly this in New Albany, Ohio. The company put tens of thousands of chips in tents next to the campus, powered by gas turbines off the grid. Each chip costs about 60,000 dollars, the report says. The first five buildings on the campus took two to three years of construction. With the tents, Meta cut in half the time it takes to switch on new chips. Companies that plan their AI projects around how fast the power company can connect them will pay a high price for a power shortage that, in my view, will last only one cycle.

Have a good day, everyone. Thirteen researchers have built a shared test for medical second opinions written by language models. It's called AI Consilium, and it came out on arXiv on October 15, 2026. They started from 300 cases in the NEJM-MedQA collection where the correct diagnosis is already known. On 110 cases, doctors gave reference scores, which were used to calibrate an automated judge. The other 190 cases were used for the real test. They tested eight systems, each built on a different model. In each system, every AI specialist wrote an opinion, and then the system wrote a final report. Every report comes in two versions. One is free-form. The other is a structured, step-by-step account of how the diagnosis was reached. The judge, another language model, scored 3,040 final reports. According to the authors, the quality of the reports depends more on the model used than on the format of the answer.

Hello, everyone. DoorDash started from what it already had in house. The delivery platform gets hundreds of thousands of help requests every day. They come from consumers, from partner stores, and from Dashers, the independent couriers. AWS reports that at the end of 2023, DoorDash had more than 37 million monthly active consumers and more than 2 million Dashers. There was already an automated voice menu built with Amazon Connect and Amazon Lex. According to AWS, that menu had cut transfers to a live agent by 49 percent. Working with the AWS Generative AI Innovation Center, DoorDash built a generative AI voice assistant on top of Amazon Connect. According to the case study AWS published, it was ready for live testing in two months, with development time cut by 50 percent, thanks to Amazon Bedrock and Anthropic's Claude model.